What happened
Klaviyo pushed two AI agents into public beta. Composer is the marketing side: it audits your live campaigns, flows, and segments, then surfaces ranked revenue opportunities — an underperforming abandoned-cart flow, a welcome journey bleeding customers, a lapsed high-value segment — and builds cross-channel campaigns to chase them. It draws on 14 years of Klaviyo data and patterns from nearly 200,000 brands.
Customer Agent is the service side, and it's the more consequential one. It comes pre-trained on your brand data, voice, and policies, and it doesn't just explain your return policy — it processes the return. It applies loyalty points. It operates across web chat, email, SMS, and WhatsApp, and can be embedded in custom surfaces. The two agents share real-time customer data with each other.
Early-access brands in the private beta included AS Beauty, Spanx, and Dermalogica. Klaviyo's CMO Jamie Domenici framed the pitch as context, arguing most AI can't act on the context that matters. Notably absent from the announcement: pricing and a general-availability date.
Why this matters
The line that separates this generation of tools from chatbots is execution. A bot that explains your return policy deflects a ticket. An agent that processes the return touches your revenue, your inventory, and your customer's trust — which means the failure mode changed too. A wrong answer used to annoy someone. A wrong action refunds the wrong order.
Email at scale amplifies whatever it's given. A bad subject line a human reviewed is one mistake. An agent generating and sending flows without review makes that mistake at list scale, at 2am, in your name. And Klaviyo's own selling point — patterns from nearly 200,000 brands — cuts both ways: that's pattern-matching toward the average, and your brand voice is precisely the thing the average erodes.
The honest read on public beta: you are the QA team, and you're doing it without knowing the eventual price. That's not a reason to skip it. It's a reason to run it the way you'd run a promising junior hire's first month — real work, low stakes, everything reviewed.
What to do about it
Start Composer on one low-stakes flow
Pick something like browse abandonment — real revenue, small blast radius. Let Composer propose and build, but route every send through human approval. Compare its output against your current flow for four weeks before you let it touch a core sequence like welcome or post-purchase.
Write Customer Agent's boundaries before it goes live
Decide in writing which actions it may execute — returns yes or no, loyalty adjustments up to what point value, refunds up to what dollar amount. Then test it against your ugliest real policy edge cases: partial returns, expired windows, final-sale items. If it improvises on any of them, it's not ready for customers.
Keep an approval gate on all outbound for the beta period
Treat autonomy as earned, not granted. Log every action both agents take, review the log weekly, and only remove the human gate on a flow after it has run clean long enough that you'd trust a new employee with the same track record.